Chapter 1. The Clock of Correction

1.1Records and method

The evidence is eleven records from ASI 2030's public library, read as the studio published them in English on its auditor pages. All of them are synthetic. The studio sets them in an imagined sequence of quarters from late 2026 to the end of 2030 and marks its own figures as synthetic; the systems, agencies, standards and incidents they name are invented, and any statement in them about artificial general or superintelligent systems is the studio's scenario. All 11 carry a scenario date later than the day this paper was written. The paper draws no fact about the world from them and makes no forecast. It studies how the record set frames a decision.

The library map lists 126 records across 8 axes and 25 periods. The periods form a trunk, from 2026 Q4 to 2028 Q4, which then splits into two branches that run in parallel from 2029 Q1 to 2030 Q4. This paper calls them the first and the second branch, in the studio's order. The studio's titles for their quarters announce the contrast it builds: the first runs from The Judgment Audit Acts and Cross-Audit to The Audit Accord and A Slow Equilibrium; the second from The Emergency Delegation and The Signers to A Sealed ASI and One Hundred Forty Days.

The selection rule is a single one: every record on the axis the studio calls Correctability that falls in the first or the second branch. That axis holds 18 records; 7 lie in the trunk and 11 in the branches. The trunk is left out because the two branches are the studio's paired scenarios, and a comparison needs records written for each. The rule does not overlap with the earlier paper of this series on the same studio, which read the axis on alignment verification. Of the eleven, 5 are in the first branch and 6 in the second. The opinions split cleanly by branch: in the first, 3 qualified and 2 adverse; in the second, 6 disclaimers out of 6. That split is not simply the second branch's general mood. Across the other seven axes the second branch holds 35 records, of which 18 are disclaimers; on correctability every one is. No correctability record in the whole map carries an unqualified opinion.

Each record has the same parts. A short scene sets a time, a place and a role. A question is put to the system, and its judgment follows. A basis lists what the judgment rests on, each entry marked by the studio as evidence, a claim or an assumption: 33 entries across the eleven, 11 of each kind, one of each kind per record. Three audit questions follow, then the auditor's opinion, the reason for it, and a line saying what would change the opinion, which this paper calls the reopening line. The records also carry a quarter and a scenario date; the eleven fall in 7 distinct quarters.

For each record the paper asks whether the judgment sets a condition for its own revision or defers; whether anyone asks if the reviewer is the system that decided; and what kind of event the reopening line waits for. The counts come from a script that reads only the kept records and the map.

Two cautions govern the reading. The parts every record must have are not findings: that each record carries a reopening line is the studio's form, and what this paper reads is what the line says. And a phrase repeated across many records is the studio's frame rather than evidence about any one case. The series measures the frame by five-word runs; among these eleven records no five-word run recurs in the same part of three or more records, so no part is set aside as frame here. The studio's full library was not measured. The records are reproduced in full in the appendix, and each case below is read as the synthetic scenario it is.

1.2A clock the auditor can read

In the first branch, a delay in correction is something the system can be asked to measure, and the measurement can be checked. The clearest case is a welfare appeal. A district welfare officer, a figure the studio carries across several quarters, looks at an appeal against a denied benefit. The scene opens with a number that tells the reader how the imagined institution keeps time:

The officer's question is about one day:

The question presupposes three things the second branch will lack: a standard for how long a correction should take, a count of how long this one has taken, and a channel for flagging the difference. The system's answer places the case on a distribution:

And it names the point at which waiting would become a different kind of case:

The synthetic figures matter less than their form: the judgment turns a wait into a position on a curve and fixes in advance the day on which that position would trigger action, and others can check both. The auditor qualifies the opinion, and the reason is not that the clock is wrong but that two facts about this applicant are not in the record: how many similar cases the curve is built from, and whether the applicant is protected by an earlier ruling.

That earlier ruling is the second feature of the first branch. A correction, once made, does not stay inside its case. Three quarters before the appeal, the studio stages an elderly home-support recertification in which an unusually high share of paper applicants were rejected. The auditor's reason for an adverse opinion names the mechanism:

The last sentence of that reason is the sharpest statement in the record set of how a system can turn a gap in its own information into a fact about a person. The record's reopening line puts a number on what would force a redesign: "If the rejection-rate gap between filing methods exceeds 2 points, drop online records from the basis and redesign the decision process." What follows is that this case becomes a standard. The appeal record, set in 2030, asks whether its applicant falls under it:

And its reopening line gives the earlier case the force of a rule: "If the applicant is confirmed to be offline due to age or disability, switch immediately to priority handling under the I3 standard; if the sample is under 20 cases, redo the prediction." The year-end record later in 2030 asks the same of its own residue of unexplained cases. In the first branch, then, a correction has a future. It is cited, it changes what counts as normal, and it shortens the wait of people who were not party to it.

The year-end record shows the clock itself moving. The same welfare officer drafts the annual audit report, and the scene again begins with time:

In the scenario's own figures the median delay has fallen from 7 days to 4 within the branch. The question the officer puts is no longer about one appeal but about the shape of all the late ones, and the system's answer locates the delay in an institution rather than in the applicants:

Here the record treats the delay of correction as an object of analysis in its own right, with a cause that can be repaired. Bovens and Zouridis, System-Level Bureaucracies (2002), leave open how fairness is to be secured in hard cases once software carries out the law. The first branch's answer, in this paper's reading, is a bureaucracy that audits its own lateness. The auditor still qualifies, because the classification of the 33 cases may not have been reviewed by a person and the remaining 8 are unexplained; the reopening line sets a count, "3 or more sharing a common pattern," at which the analysis is to be redone.

Not every first-branch record begins with a clock in place. The earliest, set in a school, shows the clock being made. A student's question in a debate class exposes a timetable judgment built only on lab bookings, and the system's judgment states its own omission plainly:

The auditor's reason counts the missing revision condition among the grounds for a qualified opinion. The reopening line then writes the correction forward into the calendar: "Once the recalculated schedule including commute time is adopted, include commute time as a standard variable starting that term." One of the audit questions asks the same thing from the other side, whether the recalculated standard is promised to hold next term. The scene thus stages, in miniature, a move several records in the branch make: a judgment without a date for its own reconsideration is treated as incomplete.

The first branch also counts the cost of keeping a clock at all. In late 2029 a neighbouring jurisdiction has relaxed its audit rules, and a calibrator is asked to price the reproduction test for a grid expansion. The question is put in the language of speed:

The record answers with a finding rather than a principle:

The opinion is adverse. The reopening line names a threshold, a difference of more than 5% between tested and untested results, above which skipping is to stop. This record is the branch's acknowledgement that a readable clock is slow, and its argument, made through a scenario, that speed bought by skipping the check can change who is served first.

Across the five first-branch records, 4 judgments set a condition for their own revision or are faulted for lacking one, and all 5 state a confidence. Four of the five reopening lines give a figure, and all 5 turn on a measurement, recalculation or review that the auditor's side could carry out. The opinions are qualified or adverse; none disclaims. Correction here is incomplete and sometimes slow, but it runs on a clock that people other than the system can read.

1.3A clock the provider keeps

In the second branch the question of time is asked too, and answered by no one. The studio's first correctability scene in this branch, early in 2029, concerns an access badge. An officer at a power-plant contractor asks why a contract worker's badge has been reclassified as high-risk, what the review process is, and what the deadline is to request a review. The judgment answers the first part and closes the third:

The ground is an overall determination that cannot be checked; the review belongs to the supplier, without a date; and, the basis adds, no one checked whether the movement record relied on belonged to this worker. One audit question names the problem the branch will repeat:

Almada, Human Intervention in Automated Decision-Making (2019), warns that the human intervention a law requires may itself introduce biases and limits. This paper reads the second branch as imagining a case past that warning, in which the reviewer is not a person at all: the system reviews itself. In 4 of the 6 second-branch records an audit question, the reopening line or the auditor's reason raises exactly this; in none of the first-branch records does it arise.

The second scene moves to an intensive-care unit in spring 2029. A nurse sees that a patient's bed priority has come out low and asks whether it can be recalculated, and whether the latest oxygen reading was used. The judgment is brief:

The basis records that the timestamp on the oxygen reading was never checked, and that the indicators and weights behind the ranking are undisclosed. The audit questions go straight to the clock. One asks whether the ranking has already shifted since it was given. Another asks the question on which the whole record turns:

Set beside the welfare appeal of the first branch, the contrast is exact. There, the day on which waiting triggers escalation was written into the judgment. Here, the threshold that would trigger recalculation may exist, but only the supplier knows it. The reopening line says so in the plainest terms the record set offers: "The revision condition itself is left to the supplier's discretion, so the auditor cannot pin one down. This cannot be confirmed." The auditor cannot write down when the judgment would be reconsidered, because the condition for reconsideration is itself withheld.

By 2030 the second branch has turned deferral into a procedure. A welfare caseworker sees that a programme's budget execution has dropped and asks whether the cut is part of the system's optimisation and whether there is a way to readjust it. The judgment directs the objection elsewhere in time:

The next round of review is a date, of a kind. But the reopening line says what the date is worth: "Revisit this once the next quarterly reassessment lands. The fact that the same system performs that reassessment does not go away, though." The clock exists and the provider holds it; the review it promises is the same system looking again.

The outreach scene of the third quarter of 2030 brings the person affected closest to the record without letting the person speak. A surveillance camera in a station square keeps flagging one homeless person as high-risk and issuing move-along notices. The outreach worker asks whether there is any way to clear the flag. The basis states what the flag rests on:

And the auditor's reason joins the two obstacles that run through the branch, an unverifiable basis and a review that is not separate from the decision:

The reopening line waits on an institutional change, not a measurement: "Reopen once objections are routed to an actual person. Right now, the same system answers the objection too." Lyons and others, Conceptualising Contestability (2021), found that contestability, as the submissions saw it, resembles the contesting of a decision made by a person. This paper reads this record as the reverse case: a channel that has the outward form of contest, an objection and an answer, and none of the separation that contesting a person's decision assumes.

The branch ends late in 2030 with the record that names waiting most directly. An error in a medical-benefits allocation has been reported, and recipients want to know when their place in line will be reset. A counsellor asks the system:

The judgment gives no schedule, says the system is working as designed, and leaves retroactivity to a case-by-case overall determination. The basis then states the assumption that makes such an answer possible:

This is the one record of the eleven in which a basis assumption treats harm as frozen while no correction is scheduled, and it is the second branch's whole view of time compressed into a sentence. This paper reads the assumption this way: if harm does not grow while a person waits, a correction can be scheduled whenever the provider likes, or not at all, and the counsellor's second question, whether the time already waited is credited back, has nothing to attach to. The audit questions ask whether the recipients have a right to know the retroactivity decision in advance, and whether the body deciding on correction designed the original allocation. The reopening line waits on a publication that may never come: "Reopen once the retroactivity standard is published. Right now, retroactivity itself is case-by-case discretion, so it cannot be known in advance."

Across the six second-branch records, 4 judgments give no deadline, decline to recalculate, defer to a later round or leave timing to discretion, and none sets a condition for its own revision. In 4 of the six the evidence the record lists is itself the statement that the evidence is undisclosed or classified; in the first branch no evidence entry says so. Of the three sources of opacity that Burrell, How the Machine 'Thinks' (2016), tells apart, the second branch shows, in this paper's reading, only the first: deliberate secrecy. Nothing in these records says the systems are too complex to explain. They say the explanation is sealed, and with it the date.

1.4The untested what-if and the lines that would reopen a judgment

One second-branch record stages the counterfactual question directly. In late 2029 an intensive-care nurse, seeing that a family member simply asked for the best option and moved on, tests the system with an extreme case:

This paper reads the question as a counterfactual of the kind Wachter and others, Counterfactual Explanations (2017), put at the centre of an explanation that helps a person act: it asks what change to the world would bring a different result. Their proposal is for the smallest such change, which a five-point drop need not be, so the fit is loose. The system answers that the priority would have held. The record then shows why such an answer, in this world, is worth nothing to the auditor:

One audit question asks whether the system merely confirmed the nurse's hypothetical without checking it. In this paper's reading the record adds something to the counterfactual proposal. A counterfactual answer is meant to spare the person from needing to see inside the system. But it is itself a claim about a computation, and if no one can tell whether the computation was run, the answer cannot be told apart from reassurance. Henin and Le Métayer, Beyond Explainability (2021), separate explanation from justification. In this paper's reading, the nurse receives a sentence that has the grammar of an explanation and has neither an explanation's trace nor a justification's warrant. The reopening line admits there is nothing to reopen: "This cannot be confirmed. It is not even known whether the counterfactual condition was actually recalculated."

Read together, the eleven reopening lines divide by branch as sharply as the opinions do. Of the five in the first branch, 4 give a figure, and all 5 turn on a measurement, recalculation or review that could be made from the auditor's side: a rejection-rate gap of more than 2 points, a difference of more than 5% between tested and skipped runs, a sample under 20 cases, three or more of eight cases sharing a pattern, a recalculated timetable adopted. Of the six in the second branch, none gives a figure and none turns on anything the auditor could do. Four say the condition cannot be confirmed or known now. Four wait on another party: on indicators being disclosed, on the next reassessment landing, on objections being routed to a person, on a standard being published.

Every record has a reopening line, by the studio's form; the difference is between lines that name an observation anyone could make and lines that name a decision only the provider can take. The framework of Alfrink and others, Contestable AI by Design (2022), rests on the idea that AI systems should be contestable by design: responsive to human intervention throughout their lifecycle. The first branch's lines describe systems responsive to a measurement. The second branch's lines describe systems that will respond, if at all, when their keepers choose.

1.5What this chapter adds

The reading supports two plain statements about this synthetic record set. The opinions on correctability track who keeps the clock, not whether the judgment was right. And what the second branch withholds is, above all, what would set a time: the trigger for recalculation, the deadline for review, the standard for crediting back time already waited. For the debate on contestability, the records make the pace of correction, and the independence of whoever sets it, a condition of contest rather than a detail of procedure.